An evaluation method and device for device security risks based on multiple neural networks

Abstract

The present invention discloses a method and device for evaluating the security risks of a device based on multiple neural networks. The method includes: obtaining measurement data of the device, and respectively using the measurement data to perform evaluation training with each neural network to construct a plurality of device security evaluation models; repeatedly performing coefficient adjustment operations on each of the device security evaluation models to generate corresponding comprehensive evaluation models; using the plurality of comprehensive evaluation models to perform security risk evaluation calculations with preset operation data, and obtaining a calculation result, where the preset operation data is status data collected in real time when the device is running; determining the device security risk based on the calculation result. The present invention can collect the measurement data of the device, and use the measurement data to perform evaluation training and weight allocation on multiple neural networks, thereby improving the evaluation accuracy of the model, and setting multiple evaluation models can adapt to more different devices, further improving the flexibility and practicality of the evaluation.

Description

Technical Field

[0001] The present invention relates to the technical field of evaluating equipment safety risks, and in particular, to a method and device for evaluating equipment safety risks based on multiple neural networks. Background Art

[0002] The power industry is closely related to people's living standards. With the development of the times and the progress of technology, more and more power equipment has been put into the power grid system, making the power grid structure increasingly large and complex. At the same time, various uncertain risks such as external force factors, operation factors, equipment factors, and management factors also affect the safe and stable operation of the power grid, increasing the operation risk of the power grid.

[0003] In order to enable each device to work safely and properly, so that the power grid operates safely and stably, the commonly used method at present is to record the operation data of each device during operation, and then use the trained model to evaluate the safety risk.

[0004] However, the currently commonly used prediction and evaluation methods have the following technical problems: Since there are many types of devices, and a single training model is difficult to adapt to multiple different devices, the evaluation and prediction results do not match the actual situation, reducing the accuracy rate. Summary of the Invention

[0005] The present invention provides a method and device for evaluating equipment safety risks based on multiple neural networks. The method can use multiple models for collaborative prediction and evaluation, which can not only adapt to different types of devices, but also improve the accuracy rate of evaluation and prediction.

[0006] The first aspect of the embodiment of the present invention provides a method for evaluating equipment safety risks based on multiple neural networks, and the method includes:

[0007] Obtain the measurement data of the equipment, and use the measurement data to perform evaluation training with each neural network respectively to construct multiple equipment safety evaluation models;

[0008] Repeat the coefficient adjustment operation for each of the equipment safety evaluation models to generate corresponding comprehensive evaluation models;

[0009] Use multiple comprehensive evaluation models to perform safety risk evaluation calculations with preset operation data to obtain calculation results, where the preset operation data is the state data collected in real time when the equipment is running;

[0010] Determine the equipment safety risk based on the calculation results.

[0011] In a possible implementation manner of the first aspect, the coefficient adjustment operation is specifically:

[0012] Determine the loss value between the prediction result of each of the device security evaluation models and the given label respectively;

[0013] Initialize and calculate the initial adjustment coefficient of each of the device security evaluation models by using the loss value between the prediction result and the given label;

[0014] Multiply the prediction result by the initial adjustment coefficient to obtain an intermediate prediction result value, calculate the support degree between the intermediate prediction result value and the prediction result, and perform normalization processing on the support degree to obtain a normalized support degree;

[0015] Multiply the normalized support degree by a preset learning rate to obtain an updated adjustment coefficient until the total loss value between the prediction result and the given label is less than a preset loss threshold.

[0016] In a possible implementation manner of the first aspect, the calculation of the initial adjustment coefficient is shown in the following formula:

[0017]

[0018] Wherein, represents the prediction result of the i-th neural network model; represents the loss value of the given label corresponding to the prediction result of the i-th neural network model;

[0019] S i represents the calculation result of the prediction result of the i-th neural network model and the loss value of the given label corresponding to the prediction result of the i-th neural network model; represents the initial adjustment coefficient of the i-th neural network model.

[0020] In a possible implementation manner of the first aspect, the calculation of the normalized support degree is shown in the following formula:

[0021]

[0022]

[0023] Wherein, is the intermediate prediction result value, Sup represents the support degree used to calculate two vectors; C i is the normalized support degree, and N represents the preset number of neural networks.

[0024] In a possible implementation manner of the first aspect, the calculation of the updated adjustment coefficient is shown in the following formula:

[0025]

[0026] Wherein, represents the update adjustment coefficient; η represents the learning rate.

[0027] In a possible implementation manner of the first aspect, η = 0.01.

[0028] In a possible implementation manner of the first aspect, the total loss value of the prediction result and the given label is calculated as shown in the following formula:

[0029]

[0030] Wherein, SSE represents a method for calculating loss; represents the set threshold;

[0031] The comprehensive evaluation model is as shown in the following formula:

[0032]

[0033] f is the comprehensive evaluation model.

[0034] In a possible implementation manner of the first aspect, the neural network includes a convolutional neural network, a fully convolutional neural network, an autoencoder, a residual network, a long short-term memory network, a gated recurrent unit, or a probabilistic neural network.

[0035] In a possible implementation manner of the first aspect, the measurement data includes data collected when the power grid equipment is operating normally and data collected when the power grid equipment is abnormal.

[0036] The second aspect of the embodiments of the present invention provides an evaluation device for equipment security risks based on multiple neural networks. The device includes:

[0037] An evaluation training module, configured to obtain measurement data of the equipment, and use the measurement data to perform evaluation training with each neural network respectively to construct a plurality of equipment security evaluation models;

[0038] A coefficient adjustment module, configured to repeatedly perform coefficient adjustment operations on each of the equipment security evaluation models to generate corresponding comprehensive evaluation models;

[0039] A calculation module, configured to use a plurality of the comprehensive evaluation models to perform security risk evaluation calculations with preset operation data to obtain a calculation result, where the preset operation data is status data collected in real time when the equipment is operating;

[0040] A determination module, configured to determine the equipment security risk based on the calculation result.

[0041] Compared with the prior art, the evaluation method and device for equipment security risks based on multiple neural networks provided by the embodiments of the present invention have the following beneficial effects: The present invention can collect the measurement data of the equipment, and use the measurement data to evaluate, train and allocate weights to multiple neural networks, thereby improving the evaluation accuracy of the model. And setting multiple evaluation models can adapt to more different devices, further improving the flexibility and practicability of the evaluation, and increasing the model functions. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 FIG. is a schematic flowchart of an evaluation method for equipment security risks based on multiple neural networks provided by an embodiment of the present invention;

[0043] Figure 2 FIG. is an operation flowchart of an evaluation method for equipment security risks based on multiple neural networks provided by an embodiment of the present invention;

[0044] Figure 3 FIG. is a schematic structural diagram of an evaluation device for equipment security risks based on multiple neural networks provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0046] The currently commonly used prediction and evaluation methods have the following technical problems: Since there are many types of equipment, and a single training model is difficult to adapt to a variety of different equipment, the evaluation and prediction results do not match the actual situation, reducing the accuracy rate.

[0047] To solve the above problems, the following specific embodiments will be used to introduce and explain in detail an evaluation method for equipment security risks based on multiple neural networks provided by the embodiments of the present application.

[0048] Refer to Figure 1 , which shows a schematic flowchart of an evaluation method for equipment security risks based on multiple neural networks provided by an embodiment of the present invention.

[0049] Among them, by way of example, the evaluation method for equipment security risks based on multiple neural networks may include:

[0050] S11. Obtain the measurement data of the equipment, and use the measurement data to respectively evaluate and train each neural network to construct a plurality of equipment security evaluation models.

[0051] In one embodiment, the measurement data includes data collected when the power grid equipment is operating normally and data collected when the power grid equipment is abnormal.

[0052] In another embodiment, the neural network includes a convolutional neural network, a fully convolutional neural network, an autoencoder, a residual network, a long short-term memory network, a gated recurrent unit, or a probabilistic neural network.

[0053] In application, the measurement data can be input into each neural network, and then each neural network is evaluated and trained to obtain a corresponding device security evaluation model.

[0054] S12. Repeatedly perform a coefficient adjustment operation on each of the device security evaluation models to generate a corresponding comprehensive evaluation model.

[0055] In actual operation, since the weights of different device security evaluation models are different, in order to improve the accuracy of the evaluation, a coefficient adjustment operation can be performed on each device security evaluation model, so that the device security evaluation models can work in coordination with each other and adapt to different devices, thereby improving the practicability and flexibility of the evaluation.

[0056] In an alternative embodiment, the coefficient adjustment operation may include the following sub-steps:

[0057] Sub-step S121. Determine the loss value between the prediction result of each device security evaluation model and the given label respectively.

[0058] Sub-step S122. Initialize and calculate the initial adjustment coefficient of each device security evaluation model by using the loss value between the prediction result and the given label.

[0059] Specifically, the calculation of the initial adjustment coefficient is shown in the following formula:

[0060]

[0061] Where, represents the prediction result of the i-th neural network model; represents the loss value of the given label corresponding to the prediction result of the i-th neural network model;

[0062] S i represents the calculation result of the loss value between the prediction result of the i-th neural network model and the given label corresponding to the prediction result of the i-th neural network model; represents the initial adjustment coefficient of the i-th neural network model.

[0063] Sub-step S123: Multiply the prediction result by the initial adjustment coefficient to obtain an intermediate prediction result value, calculate the support degree between the intermediate prediction result value and the prediction result, and perform normalization processing on the support degree to obtain a normalized support degree.

[0064] Specifically, the calculation of the normalized support degree is shown in the following formula:

[0065]

[0066] Where, is the intermediate prediction result value, Sup represents the support degree used to calculate two vectors; C i is the normalized support degree, and N represents the number of preset neural networks.

[0067] Sub-step S124: Multiply the normalized support degree by the preset learning rate to obtain an updated adjustment coefficient until the sum of the loss values of the prediction result and the given label is less than the preset loss threshold.

[0068] Specifically, the calculation of the updated adjustment coefficient is shown in the following formula:

[0069]

[0070] Where, represents the updated adjustment coefficient; η represents the learning rate.

[0071] Optionally, η = 0.01.

[0072] In one of the embodiments, the sum of the loss values of the prediction result and the given label is calculated as shown in the following formula:

[0073]

[0074] Where, SSE represents a method for calculating loss; represents the set threshold;

[0075] The comprehensive evaluation model is shown in the following formula:

[0076]

[0077] f is the comprehensive evaluation model.

[0078] In actual operation, the loss between the prediction result of each network model and the given label can be calculated first, and the initial adjustment coefficient of each network model can be initialized; the prediction result of each network is multiplied by the corresponding network adjustment coefficient to obtain an intermediate prediction result value, the support degree of this intermediate value and the network prediction result is calculated, and the support degree is normalized to obtain a normalized support degree; the normalized support degree is multiplied by a preset learning rate to obtain an updated adjustment coefficient; then the above steps are repeated, and the updated adjustment coefficient is multiplied by the prediction result of each network respectively to obtain an intermediate prediction result value until the comprehensive evaluation value of the device security risk and the loss of the given label drop to a certain threshold and stop, to obtain the target adjustment coefficient, and then the final comprehensive evaluation model of the device security is determined according to the target adjustment coefficient, and the final comprehensive evaluation model of the device security is used as the comprehensive evaluation model.

[0079] S13. Use multiple said comprehensive evaluation models to perform security risk assessment calculations with preset operation data to obtain calculation results, wherein the preset operation data is status data collected in real time when the device is running.

[0080] When the device is running, the status data of the device can be collected in real time, and the status data can include data on the device status such as the running time, running frequency, power, signal transmission path, etc. of the device. Finally, the status data can be input into the comprehensive evaluation model, and the corresponding security risk assessment calculation is performed by the comprehensive evaluation model.

[0081] In an optional embodiment, the operation data can be collected synchronously when collecting measurement data to improve the data collection efficiency.

[0082] S14. Determine the device security risk based on the calculation results.

[0083] Finally, the device security risk can be determined based on the calculation results.

[0084] Optionally, the device security risk can be determined according to the magnitude of the calculated value. For example, the higher the value, the greater the risk, and the lower the value, the lower the risk.

[0085] Refer to Figure 2 , which shows an operation flowchart of a method for evaluating the device security risk based on multiple neural networks provided by an embodiment of the present invention.

[0086] Specifically, first obtain the measurement data of the device and the operation data during operation, then use the measurement data to evaluate and train multiple neural networks, train each neural network into a corresponding comprehensive device security evaluation model, combine multiple comprehensive device security evaluation models and adjust the weights of each comprehensive device security evaluation model to generate a corresponding comprehensive evaluation model, and finally input the operation data into the comprehensive evaluation model for risk assessment calculation to determine the security risk of the device.

[0087] In this embodiment, the embodiment of the present invention provides a method for evaluating the security risk of a device based on multiple neural networks, and its beneficial effects are as follows: The present invention can collect the measurement data of the device, and use the measurement data to evaluate, train and assign weights to multiple neural networks, thereby improving the evaluation accuracy of the model. And setting multiple evaluation models can adapt to more different devices, further improving the flexibility and practicality of the evaluation, and increasing the model functions.

[0088] The embodiment of the present invention also provides an apparatus for evaluating the security risk of a device based on multiple neural networks. Refer to Figure 3 , which shows a schematic structural diagram of an apparatus for evaluating the security risk of a device based on multiple neural networks provided by an embodiment of the present invention.

[0089] Among them, by way of example, the apparatus for evaluating the security risk of a device based on multiple neural networks may include:

[0090] An evaluation and training module 301, configured to obtain the measurement data of the device, and use the measurement data to respectively evaluate and train each neural network to construct multiple device security evaluation models;

[0091] A coefficient adjustment module 302, configured to repeatedly perform coefficient adjustment operations on each of the device security evaluation models to generate corresponding comprehensive evaluation models;

[0092] A calculation module 303, configured to perform security risk assessment calculations on multiple comprehensive evaluation models and preset operation data to obtain a calculation result, where the preset operation data is status data collected in real time during the operation of the device;

[0093] A determination module 304, configured to determine the security risk of the device based on the calculation result.

[0094] Optionally, the coefficient adjustment module is further configured to:

[0095] Respectively determine the loss values of the prediction results of each of the device security evaluation models and the given labels;

[0096] Initialize and calculate the initial adjustment coefficient of each of the device security evaluation models by using the prediction result and the loss value of the given label;

[0097] Multiply the prediction result by the initial adjustment coefficient to obtain an intermediate prediction result value, calculate the support degree of the intermediate prediction result value and the prediction result, and perform normalization processing on the support degree to obtain a normalized support degree;

[0098] Multiply the normalized support degree by a preset learning rate to obtain an updated adjustment coefficient until the total loss value of the prediction result and the loss value of the given label is less than a preset loss threshold.

[0099] Optionally, the calculation of the initial adjustment coefficient is shown in the following formula:

[0100]

[0101] where, represents the prediction result of the i-th neural network model; represents the loss value of the given label corresponding to the prediction result of the i-th neural network model;

[0102] S i represents the calculation result of the prediction result of the i-th neural network model and the loss value of the given label corresponding to the prediction result of the i-th neural network model; represents the initial adjustment coefficient of the i-th neural network model.

[0103] Optionally, the calculation of the normalized support degree is shown in the following formula:

[0104]

[0105] where, is the intermediate prediction result value, Sup represents the support degree used to calculate two vectors; C i is the normalized support degree, and N represents the preset number of neural networks.

[0106] Optionally, the calculation of the updated adjustment coefficient is shown in the following formula:

[0107]

[0108] where, represents the updated adjustment coefficient; η represents the learning rate.

[0109] Optionally, η = 0.01.

[0110] Optionally, optionally, the coefficient adjustment module is further configured to:

[0111]

[0112] Among them, SSE represents a method for calculating loss; represents setting a threshold;

[0113] The comprehensive evaluation model is shown as follows:

[0114]

[0115] f is the comprehensive evaluation model.

[0116] Optionally, the neural network includes a convolutional neural network, a fully convolutional neural network, an autoencoder, a residual network, a long short-term memory network, a gated recurrent unit, or a probabilistic neural network.

[0117] Optionally, the measurement data includes data collected when the power grid equipment is operating normally and data collected when the power grid equipment is abnormal.

[0118] Furthermore, an embodiment of the present application also provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the method for evaluating the device security risk based on multiple neural networks as described in the above embodiment.

[0119] Furthermore, an embodiment of the present application also provides a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to cause a computer to execute the method for evaluating the device security risk based on multiple neural networks as described in the above embodiment.

[0120] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A method for evaluating the security risk of a device based on multiple neural networks, characterized in that, the method includes: Obtain the measurement data of the device, and use the measurement data to perform evaluation training with each neural network respectively to construct multiple device security evaluation models; Repeat the coefficient adjustment operation for each of the device security evaluation models to generate corresponding comprehensive evaluation models; Use multiple of the comprehensive evaluation models to perform security risk evaluation calculations with preset operating data to obtain calculation results, where the preset operating data is the status data collected in real time when the device is running; Determine the device security risk based on the calculation results; The coefficient adjustment operation is specifically: Respectively determine the loss values of the prediction results of each of the device security evaluation models and the given labels; Use the prediction results and the loss values of the given labels to initialize and calculate the initial adjustment coefficients of each of the device security evaluation models; Multiply the prediction results by the initial adjustment coefficients to obtain intermediate prediction results, calculate the support degrees of the intermediate prediction results and the prediction results, and perform normalization processing on the support degrees to obtain normalized support degrees; Multiply the normalized support degrees by a preset learning rate to obtain updated adjustment coefficients until the total loss value of the prediction results and the loss values of the given labels is less than a preset loss threshold.

2. The method for evaluating the security risk of a device based on multiple neural networks according to claim 1, characterized in that, The calculation of the initial adjustment coefficient is shown in the following formula: ； ； Among them, represents the prediction result of the i-th neural network model; represents the loss value of the given label corresponding to the prediction result of the i-th neural network model; Indicates the calculation result of the loss value of the given label corresponding to the prediction result of the th neural network model and the prediction result of the Indicates the initial adjustment coefficient of the 3. The method for evaluating the security risk of a device based on multiple neural networks according to claim 2, characterized in that, The calculation of the normalized support degree is shown in the following formula: ； ； ； Among them, is the intermediate value of the prediction result, indicating the support degree used to calculate two vectors; is the normalized support degree, indicating the number of preset neural networks.

4. The method for evaluating the security risk of a device based on multiple neural networks according to claim 3, characterized in that, The calculation of the updated adjustment coefficient is shown in the following formula: ； Among them, represents the update adjustment coefficient; represents the learning rate.

5. The method for evaluating the security risk of a device based on multiple neural networks according to claim 4, characterized in that, The said = 0.

01.

6. The method for evaluating the security risk of a device based on multiple neural networks according to claim 4, characterized in that, The calculation of the total loss of the prediction results and the loss values of the given labels is shown in the following formula: ； Among them, represents a method for calculating loss; represents a set threshold value; The comprehensive evaluation model is shown in the following formula: ； It is a comprehensive evaluation model.

7. The method for evaluating the security risk of a device based on multiple neural networks according to any one of claims 1-6, characterized in that, The neural network includes a convolutional neural network, a fully convolutional neural network, an autoencoder, a residual network, a long short-term memory network, a gated recurrent unit or a probabilistic neural network.

8. The method for evaluating the security risk of a device based on multiple neural networks according to any one of claims 1-6, characterized in that, The measurement data includes the data collected when the power grid device is operating normally and the data collected when the power grid device is abnormal.

9. An apparatus for evaluating the security risk of a device based on multiple neural networks, characterized in that, the apparatus includes: An evaluation and training module, configured to obtain measurement data of a device, and respectively perform evaluation and training on each neural network by using the measurement data to construct a plurality of device security evaluation models; A coefficient adjustment module, configured to repeatedly perform coefficient adjustment operations on each of the device security evaluation models to generate corresponding comprehensive evaluation models; A calculation module, configured to perform security risk assessment calculations on a plurality of the comprehensive evaluation models and preset operation data to obtain calculation results, wherein the preset operation data is status data collected in real time when the device is operating; A determination module, configured to determine the device security risk based on the calculation results; The coefficient adjustment operation specifically is: respectively determine the loss values between the prediction results of each of the device security evaluation models and a given label; initialize and calculate the initial adjustment coefficients of each of the device security evaluation models by using the prediction results and the loss values of the given label; multiply the prediction results by the initial adjustment coefficients to obtain intermediate prediction result values, calculate the support degrees between the intermediate prediction result values and the prediction results, and perform normalization processing on the support degrees to obtain normalized support degrees; multiply the normalized support degrees by a preset learning rate to obtain updated adjustment coefficients until the total loss value between the prediction results and the loss values of the given label is less than a preset loss threshold.

Images